• 9:30am - Jenna DeCata
  • "Multiple isoforms of RNA-binding protein Bruno1 are required during indirect flight muscle development in Drosophila"
  • Abstract:
    • Indirect flight muscles (IFMs) of Drosophila undergo a switch in fiber-type specific alternative splicing that is essential for flight behavior. This switch is regulated by RNA-binding proteins (RBPs), especially Bruno1 (Bru1). Bru1 is a conserved member of the CELF family, which are important regulators of developmental alternative splicing in vertebrate striated muscle. CELF activity is misregulated in patients with Myotonic Dystrophy Type I, notably reverting to embryonic splicing patterns. In flies, Bru1 has a role in translation repression during embryo formation. Recently, Bru1was identified as a muscle fiber-type specific splicing factor necessary for a developmental transition to IFM-specific splice isoforms. Bru1 promotes early cytoskeletal rearrangements enabling myofibrillogenesis and maturation of the sarcomeres and contractility. Bru1 is alternatively spliced, but isoform-specific CELF functions are not reported in flies or vertebrates. Here we test if Bru1 isoforms play distinct roles in myogenesis. We show that an isoform-specific CRISPR mutant affecting long Bru1 isoforms develops normally but has a hypercontraction phenotype. Using Gal4-UAS to control when and where specific Bru1 isoforms are expressed, we test the function of long (isoform B), middle (isoform A), and short (isoform D) isoforms. Our results show distinct localization patterns for Bru1 isoforms in myogenesis. We show that overexpression of Bru1-isoB results in a strong muscle detachment phenotype, while Bru1-isoA overexpression disrupted sarcomere structure. Bru1-isoD overexpression resulted in only minor muscle defects. We further tested isoform-specific rescue ability and find that a combination of Bru1-isoA and Bru1-isoD are required to rescue IFM development. These differences suggest a requirement for distinct Bru1 isoforms during muscle development and are the first demonstration of CELF-family isoform-specific function in myogenesis.
• 9:50am - Alyssa Stanfield
  
  • "Engineering biomolecular condensates for limiting Tobacco mosaic virus accumulation and enhancing disease resistance"
  • Abstract:
    • Artificial condensates formed by liquid-liquid phase separation (LLPS) have been extensively studied in the field of synthetic biology for their ability to control cellular functions in a reversible and tunable manner. While several naturally occurring cellular condensates can restrict viral accumulation, the potential to employ LLPS as a platform for targeted antiviral strategies remains understudied. We engineered Tobacco mosaic virus (TMV) to harbor MS2 bacteriophage hairpins that are specifically bound by the MS2 coat protein (MCP). Using well-studied intrinsically disordered regions as scaffolds, we created artificial condensates that could target MS2-tagged TMV and inhibit TMV accumulation up to 5-fold in transient assays in Nicotiana benthamiana. Furthermore, plants expressing artificial condensates designed to sequester MS2-tagged TMV exhibited >10-fold reduction in systemic TMV trafficking and significantly reduced symptom severity following mechanical inoculation of purified virions. This study shows that LLPS-based antiviral strategies can effectively target viral RNAs, suppress virus accumulation, and has the potential to aid efforts to minimize crop loss.
• 10:10am - Nafisa Tasnam
  • "Direct Carbon Electrode Coating Enabled by a Simple Organic Hole-Transporting Material for Fully Solution-Processed Perovskite Solar Cells"
  • Abstract:
    • Hole transporting materials (HTMs) in perovskite solar cells play vital roles in device performance and the ease of device fabrication. We have previously shown that a substituted diacenaphtho[1,2-b:1′,2′-d]thiophene (DAT) is an excellent HTM for perovskite solar cells as it is hydrophobic, has matching frontier orbitals to MAPbI3, and has strong thin film hole mobility without doping. We demonstrate in this presentation that DAT creates dense, pinhole-free films on top of perovskite, which enables wet carbon paste to be directly coated by doctor-blading to create a top carbon electrode that is not possible with the most popular HTM Spiro-OMeTAD film. Carbon electrodes offer a simple, low-cost, and eco-friendly substitute to vacuum-deposited Au/Ag. A fully solution-based approach has been used to construct devices up to 1 cm2 in size with the configuration of FTO/dense-TiO2/mesoporous-TiO2/MAPbI3/ (with or without) DAT/Carbon in ambient conditions. In the absence of the DAT, the best device with a directly coated carbon electrode produced a Voc of 0.92 V, Jsc of 22.43 mA/cm2, a fill factor (FF) of 0.45, and a power conversion efficiency (PCE) of 9.2%. After only 5 days of ambient storage, these values decreased to 0.73 V for Voc, 11.95 mA/cm2 for Jsc, 0.35 for FF, and 3.0% for PCE, respectively. The device with the DAT, on the other hand, demonstrated significantly better performance, exhibiting Voc of 1.02 V, Jsc of 22.93 mA/cm2, FF of 0.69, and PCE of 16.1%. The device achieved one of the highest efficiencies for MAPbI₃/carbon-based solar cells, only slightly lower than similar devices with Au electrodes [(17.23 ± 0.71)%]. The DAT and carbon layers enabled the device to maintain over 91% efficiency after three months at room temperature without any encapsulation. It performs better than both DAT/Au and carbon-only devices without the DAT layer. We are working on new HTM-diacenaphtho[1,2-b:1′,2′-d]pyrrol (DAP) to fabricate higher-efficiency perovskite solar cells.
• 10:30am - Syfur Tushar
  • "Porous Crosslinked Polyethyleneimine Networks for Selective and Fast Gold Adsorption from Electronic Waste"
  • Abstract:
    • Electronic waste, or e-waste, is one of the fastest-growing waste streams as electronic gadgets become increasingly prevalent and short-lived. E-waste contains a significant amount of precious metal gold, whose retrieval bears substantial financial and environmental benefits. The challenge is to develop an adsorbent that can capture Au with high capacity, high selectivity, and fast rates. Triazine crosslinked polyethyleneimines (TCPEIs) were previously reported to potentially be such adsorbents. To investigate whether the adsorption capacity and rates, particularly at low Au concentrations, can be further improved, we studied crosslinked polyethyleneimine (CPEIs) synthesized with rigid crosslinkers of varying lengths and different PEI types. Specifically, a triazine (2,4,6-trichloro triazine (TCT)) and a piperazine-bridged triazine dimer (1,4-di(4,6-dichloro-1,3,5-triazine-2-yl) piperazine (DTP)) were used to crosslink branched or linear PEIs to generate four CPEIs – bTCPEI, bDCPEI, lTCPEI, lDCPEI. Langmuir isotherm analysis revealed maximum adsorption capacities of 2941, 2564, 2857, and 2778 mg/g for bTCPEI, bDCPEI, lTCPEI, and lDCPEI, respectively, indicating that neither PEI type nor crosslinker size significantly affects the maximum adsorption capacity. lDCPEI consistently showed faster adsorption kinetics across all three tested initial gold concentrations than lTCPEI. At 20 ppm, lDCPEI exhibited a rate constant of 0.5837 g·mg⁻¹·min⁻¹ and 94% removal efficiency, versus 0.2231 g·mg⁻¹·min⁻¹ and 77% for lTCPEI. XRD analysis indicated that the predominant mechanism of initial adsorption is electrostatic interactions, followed by the reduction of gold ions. Even with a 50-fold surplus of competing metals, lDCPEI was able to recover 99% of the gold. Moreover, the desorption study showed that 99% of the adsorbed gold could be recovered within 30 seconds.This simple approach to e-waste recycling offers sustainable benefits in the economy and environmental protection.

Room 238:

• 9:30am - Emily Trapani
  • "Navigating Neoliberalism \ Eighth Blackbird"
  • Abstract:
    • This research aims to illuminate the nuanced market dynamics that 21st century classical musicians must adhere to in order to cultivate a successful career through the analysis of the shifting programming and artistic decisions of the contemporary chamber music ensemble, Eight Blackbird across time.
      
      As we enter the first quarter of the twenty-first century, waning public interest in classical music has required industry groups, musicians, and advocacy organizations to shift their business model to resemble that of a private corporation competing in a competitive marketplace. Their establishment of niche communities within the marketplace are most observable through ensemble branding, programming, and marketing tactics; these realities are a byproduct of classical musicians’ adaptation to a mercurial neoliberal global climate that has come to define the economic landscape of the past thirty years. Through the exploration of the semiotic meanings underlying the programmatic choices of contemporary chamber ensemble Eighth Blackbird, this paper seeks to illuminate the ways in which this ensemble has successfully navigated an expanding neoliberal marketplace by embracing the principles within which it operates. In the larger conversation regarding the relevance of classical music and future of it from the stance of a performer, Eighth Blackbird’s aesthetic and programmatic choices serves as an example of what is demanded to successfully cultivate a career as a 21st century musician: adherence to the principles of a neoliberal marketplace, including their establishment of niche communities and, entrepreneurially- minded branding decisions.
• 9:50am - Melody Brooks
  
  • "The Importance of Arioso Signing in Early Childhood Music Education"
  • Abstract:
    • Arioso singing, or spontaneous vocal improvisation, is vital for developing young children's musicality. Coined by John Feierabend, this "babble singing" represents children's natural melodic experimentation, mirroring speech development. Research shows arioso strengthens pitch awareness, rhythmic skills, and creative confidence while supporting cognitive and linguistic growth.
      
      Practical classroom applications include: story-based melodic improvisation, puppet-assisted singing to reduce shyness, guided vocal play using graphic notation, and call-and-response exercises.
      
      These methods nurture innate musicality while building foundational skills for formal music education. By encouraging spontaneous singing, teachers foster creativity, self-expression, and lifelong musical engagement in young learners.
      
      This presentation aims to guide educators in utilizing arioso in their lesson to help strengthen students' musical abilities as well as singing confidence.
• 10:10am - Ian Chung
  • "Composing Democracy: Political Music of Isang Yun and the Construction of Collective Identity"
  • Abstract:
    • This paper examines how Isang Yun utilized his music as a form of public defiance against authoritarian repression, solidifying his role as a cultural symbol of resistance. Yun’s compositions and international reputation amplified South Korea’s democratic movement, demonstrating the transformative power of art in advocating for human rights and political freedom. The paper begins with an overview of South Korea’s democratic movement, focusing on the 1980 Gwangju Uprising and self-immolation as a form of political protest. Then, it transitions to an examination of how these social contexts influenced Yun to compose, Exemplum in memoriam Kwangju (1981) and Engel in Flammen mit Epilog (1994) reflecting South Korea’s struggle for democracy. Finally, using Alberto Melucci’s model of collective identity—developed from studies on 1980s social movements—the analysis examines three key aspects: (1) the use of music as a medium of resistance, (2) the fostering of active relationships among participants, and (3) the construction of a shared identity through mutual recognition. The paper concludes with how Yun’s music actively builds a collective identity and advances commitment to the goals of democratization in South Korea.
• 10:30am - David Seaton
  • "Understanding the Breakdown of the Accord Between Capital and Labor as well as the Disintegration of the American Ruling Class in Through a Synthesis of Marxist and Institutionalist Political Economy"
  • Abstract:
    • From the 1930s to the 1970s, the American ruling class entered into what has become known as the accord between capital and labor. The American ruling class was largely conciliatory with the working class. Frightened by both the impact of the Great Depression and the specter of Communism, the American ruling class saw the accord between capital and labor as giving up a little to save a lot. From the stagflation crisis of the 1970s onward, the American ruling class abandoned this conciliatory strategy in favor of one designed to claw back any concessions that they had made in the preceding decades. The American ruling class no longer saw the accord between capital and labor as giving up a little to save a lot but rather giving away the store; hence the abandonment of the strategy. Both Marxist and Institutionalist political economy both provide an antidote to understanding historical responses to economic crisis through the Great Man Theory: e.g. whether Franklin Delano Roosevelt was a good guy or whether Jimmy Carter and Ronald Regan were bad guys in light of their respective responses to the Great Depression and the Stagflation Crisis of the 1970s. Yet, Marxism focuses predominantly on class potentially to the detriment of understanding the impact of other social forces; while Institutionalism focuses on the impact of other social forces potentially to the detriment of understanding the impact of class. It is through the synthesis of Marxist and Institutional Political Economy that the full picture can be seen.

• 1:30pm - Samuel Akinyede
  • "An Enhanced Hybrid YOLOv8-MobileViT Approach for Real-Time Apple Variety Classification and Ripeness Detection"
  • Abstract:
    • We introduce a novel hybrid deep learning framework designed for real-time apple detection, variety classification, and ripeness assessment in orchard environments. This framework integrates YOLOv8 for efficient object detection with MobileViT for fine-grained classification tasks, offering a unified solution to challenges such as partial occlusions and Kaolin clay coatings on apples. To enhance performance, both YOLOv8 and MobileViT are augmented with advanced attention mechanisms, including Efficient Channel Attention (ECA), Flow Attention, and Adaptive Attention modules. These additions enable the model to better capture subtle visual cues critical for accurate detection and classification in complex orchard settings.
      
      Our work is supported by a large-scale dataset comprising over 65,000 images of nine apple varieties, with 85% of samples featuring Kaolin clay coatings—a common agricultural practice that often confounds traditional detection systems. Experimental results demonstrate that the proposed pipeline achieves state-of-the-art performance, with an mAP of 88.7% at IoU=0.5 for detection tasks and 86.2% when incorporating ripeness classification. Compared to baseline YOLOv8, our approach delivers a 3% improvement in accuracy while maintaining real-time inference speeds of 30–35 FPS on mid-range GPUs, ensuring practicality for large-scale orchard operations. Further analyses, including confusion matrices and Grad-CAM visualizations, confirm the effectiveness of attention mechanisms in improving localization and classification of partially occluded or clay-coated apples. These findings underscore the robustness of our framework in addressing real-world agricultural challenges.
• 1:50pm - Saeed Alqarni
  
  • "SAM-KG: Knowledge Graph-Enhanced Tumor Segmentation with Zero-Shot Foundation Models"
  • Abstract:
    • Accurate tumor segmentation in medical imaging is challenging due to anatomical variability, tumor heterogeneity, and data scarcity. While deep learning models like UNet, Swin-UNETR, and D-LKA Net achieve strong results, they require extensive labeled data and struggle with generalization. Zero-shot models like SAM and MedSAM offer promising alternatives but lack volumetric consistency and structured spatial reasoning. We propose SAM-KG, a knowledge graph-enhanced segmentation framework that integrates multi-modal feature extraction, uncertainty-guided refinement, and anatomical relationship modeling to improve segmentation accuracy and interpretability. Evaluated on the Liver Tumor Segmentation (LiTS) and Medical Segmentation Decathlon (MSD) Pancreas datasets, SAM-KG surpasses state-of-the-art models, achieving superior boundary delineation and volumetric consistency, advancing AI-driven medical image analysis.
• 2:10pm - Luke Miller
  • "Line Graph Constructions for Hybrid Quantum Learning on Edge-Centric Graphs"
  • Abstract:
    • We evaluate hybrid classical-quantum encoding strategies for learning on edge-featured graphs, using synthetic power network topologies where edge weights reflect critical infrastructure. The task is to predict the edge whose removal is most disruptive. To represent edge attributes in a node-centric format suitable for quantum processing, we construct the line graph—a graph where each node represents an edge in the original—and simulate a quantum circuit in Qiskit to encode node features as statevectors for downstream GNN training.
      We benchmark two line-graph encoding variants: register-based binary encoding and amplitude-based rotation, each with pairwise or full entanglement. In parallel, we test two encodings on the original graph: edgewise ancilla rotations and parameterized entanglement gates scaled by edge weights.
      We hypothesize that encoding the most information-dense elements—edges—into quantum states can help retain task-relevant structure. Though line graphs expand circuit width, they enhance access to path and connectivity patterns. Early results on small synthetic datasets with graph convolutional networks show increased performance on smaller datasets. This initial study provides a controlled testbed for evaluating quantum encodings in hybrid pipelines and may inform future work in modeling infrastructure networks or extending to fully quantum learning architectures.
• 2:30pm - Chuheng Xiao
  • "Grocery Tracker and Recommendation Generator AI for Smarter Grocery Choices and Meal Planning"
  • Abstract:
    • Grocery Tracker & Recommendation Generator is driven by a clear mission: to reduce food waste, improve meal planning, and empower users to make healthier, more efficient grocery decisions with minimal effort. In a world where busy lifestyles often lead to poor food organization and uninformed choices, our goal is to create an intelligent, user-friendly system that bridges the gap between what people have and what they could do with it.
      At the heart of the app is a personalized experience. Users can track their groceries effortlessly by simply uploading photos, receive AI-powered meal and grocery suggestions tailored to their health goals and dietary needs, and generate recipes based solely on what’s in their pantry. Whether the user is focused on saving time, reducing waste, or eating better, our app adapts to their needs.
      Unlike existing tools that treat inventory, planning, and cooking as separate workflows, we unify them. The system leverages computer vision and natural language processing in a single, seamless Streamlit interface. Google’s Gemini 1.5 model powers image recognition, while a recommendation engine curates smart suggestions. The platform is designed with future enhancements in mind, such as visual recipe instructions, nutritional insights, mobile optimization, and scalable backend support via MongoDB.
      By blending personalization, simplicity, and smart automation, Grocery Tracker & Recommendation Generator redefines how people engage with their food—from pantry to plate.

Room 238:

• 1:30pm - Udiptaman Das
  • "KARMA: Turning Healthcare Data into Actionable Knowledge with AI"
  • Abstract:
    • The growing complexity and scale of healthcare data—from structured clinical records to unstructured notes and genomics—pose significant challenges for effective analysis, integration, and clinical decision-making. While AI models such as transformers and biomedical language models (e.g., BioBERT, SciBERT) have enhanced diagnostics and predictive capabilities, they often lack explainability, semantic structure, and interoperability with clinical standards.
      
      We present KARMA (Knowledge Application for Reports in Medical Analysis), a novel framework that leverages large language models (LLMs) and medical ontologies to construct patient-centric, ontology-enriched knowledge graphs. Using Gemini Flash 2.0, each step of the KG pipeline—from entity and attribute extraction to schema generation and RDF structuring—is LLM-guided. Standard ontologies like SNOMED CT and ICD-11 are integrated for semantic alignment, enabling scalable and interoperable graph construction.
      
      KARMA further incorporates AI-driven community detection (Louvain modularity) to identify clinically meaningful patient subgroups and applies semantic reasoning (SPARQL, SWRL) to support personalized recommendations. A dynamic query system enables clinicians to explore graphs interactively using natural language inputs.
      
      To enhance graph quality, a second-pass refinement using GPT-o1 and Grok restructures the entire KG by normalizing entity-attribute-sub attribute hierarchies and refining relationships and schema. This step improves semantic precision, reduces hallucinations, and increases trust in LLM-driven clinical reasoning. Evaluation will include metrics such as ontology coverage, semantic consistency, hallucination rate, and information gain.
• 1:50pm - Gunjeet Kaur
  
  • "Silk Fibroin in Biosensing: Innovation in Implantable and Wearable Technologies"
  • Abstract:
    • Recent advancements in implantable biosensors highlight the increasing adoption of natural fiber-based materials due to their superior biocompatibility, degradable and mechanical performance. silk fibroin (SF), a natural protein fiber, has emerged as an exceptional candidate for next generation biosensors offering sustainable flexible and biodegradable solutions. its inherent mechanical strength, moisture responsiveness and tunable degradation rates make it ideal for implantable sensing platforms. SF’s capacity to form conductive hydrogels and film with efficient ionic transport enhances the development of flexible humidity sensors with rapid response time, enabling full time physiological monitoring. Moreover, silk fibroin-based nanofiber structures incorporated in wound dressing promote tissue regeneration while maintaining mechanical stability accelerating healing processes silk fibroin films have demonstrated increased sensitivity in ionic touchscreen providing multimodal sensing capabilities. silk fibroin (SF’s)substrate integrated with liquid crystal elastomers based flexible perovskite solar cells addresses thermal stress challenges while enhancing mechanical robustness. The incorporation of conductive nanomaterials like MXene , polyaniline (PANI) and carbon nanotubes significantly enhance silk fibroin electrical and mechanical properties enabling the development of sensitive strain sensors. These advancements established silk fibroin as a versatile platform for high performance implantable biosensors ensuring superior biocompatibility reduced inflammatory response and seamless integration for real time health monitoring and advanced therapeutic application.
• 2:10pm - M. Atikur Rahman
  • "Controlling Growth of 2D Perovskite Through Novel Spacer Cation Engineering for Optimized Sensing"
  • No abstract provided

• S1 - Mohammad Abdel Jawad
  • with Ricardo Moniz 
  • "Temporal Trends in Post-Resuscitation Fever After In-Hospital Cardiac Arrest"
  • Abstract:
    • Background:
      A goal of post-resuscitation care among patients successfully resuscitated from in-hospital cardiac arrest (IHCA) is avoidance of fever. However, the incidence of post-resuscitation fever after the initial therapeutic hypothermia trials in 2002 and after the recent Targeted Temperature Management (TTM) trial in 2013 is unknown.
      
      Objective:
      Examine temporal trends in fever during the first 24 hours after return of spontaneous circulation (ROSC) from IHCA during 2005-2013 (after the initial hypothermia trials) and then during 2014-2022 (after the TTM trial).
      
      Methods:
      Within the Get With The Guidelines-Resuscitation registry for IHCA in the U.S., we identified adult patients with ROSC after an index IHCA from 127 hospitals that submitted data on IHCA during both time periods between 2005 and 2022. Patients with sepsis and COVID-19 infection were excluded. We evaluated temporal trends in post-resuscitation fever (defined as >100 °F) during 2005-2013 after the initial hypothermia trials, and then between 2014-2022 after the TTM trial.
      
      Results:
      Among 41,155 patients with ROSC after IHCA, the mean age was 64.8 years (±15.0); 60.0% were male, and 68.6% were of White race. Overall, 11,745 (28.5%) developed post-resuscitation fever (Figure 1). Following the therapeutic hypothermia trials in late 2002, the incidence of fever decreased from 39.1% in 2005 to 29.0% in 2013 (P for trend < 0.001) (Figure 2). After the publication of the TTM trial in late 2013, post-resuscitation fever in the years 2014-2022 did not go up but declined more modestly (P for trend = 0.003).
      
      Conclusions:
      Between 2005 and 2013, the incidence of post-resuscitation fever after IHCA decreased substantially. Since the publication of the TTM trial in late 2013, fever incidence has not increased; rather, it has remained relatively stable, even as reported use of therapeutic hypothermia has declined
• S2 - Mohammed Alhashem
  • with Srikanth Yelem, Masud Chowdhury, Ahmed Hassan 
  • "Fuel Efficiency and Economic Assessment of a Hybrid Power Supply System for Mission Critical Applications"
  • Abstract:
    • Hybrid power supplies leveraging renewable energy sources have emerged as pivotal solutions ensuring uninterrupted power for critical applications like telecom towers in remote regions.This paper evaluates the feasibility and efficacy of a hybrid power supply integrating a liquid propane (LP) generator, Battery Energy Storage (BES), and Photovoltaic Panel (PV). The study investigates the fuel efficiency of the LP generator with a hybrid system incorporating BES and BES combined with PV across diverse loading conditions and geographical locations such as Kansas City, Ontario, and Nigeria. Furthermore, it conducts a comparative analysis of the total cost of ownership between a direct LP system and hybrid systems equipped with BES and PV panels. Simulation results conducted via an Excel spreadsheet underscore that the hybrid PV system achieves superior fuel efficiency and monthly fuel savings when the LP generator operates at 80% rated capacity compared to 60% rated capacity. During a 20-year span, the hybrid PV system with the LP operating at 80% rated capacity produces significant savings compared to the direct configuration of the LP generator.
• S3 - Md Ruhul Amin
  • with Nathan Oyler, Michelle Paquette
  • "Investigating Next Generation Thermally Conductive Materials such as Porous Materials in Semiconducting Application"
  • Abstract:
    • Efficient thermal management is critical for the performance and reliability of next-generation semiconductor devices. Covalent organic frameworks (COFs), particularly two-dimensional (2D) COFs, have emerged as promising candidates for thermally conductive porous materials due to their highly ordered structures, tunable porosity, and exceptional thermal transport properties. This study will investigate the potential of 2D COFs for semiconductor applications, focusing on their thermal conductivity mechanisms, phonon transport behavior, and integration feasibility. We will explore the impact of structural modifications, functionalization strategies, and interlayer interactions on their thermal and electronic performance. Additionally, we will evaluate the compatibility of 2D COFs with semiconductor fabrication processes and their stability under operational conditions. Through a combination of experimental approaches, this research aims to optimize 2D COFs for heat dissipation in electronic devices, providing insights into their role in advancing semiconductor technology.
• S4 - Elham Basunduwah
  • with Baek-Young Choi
  • "State-of-the-Art Crowd Counting Methods Under Complex Occlusion Patterns"
  • Abstract:
    • Crowd counting in complex scenarios with various occlusion
      patterns remains a significant challenge in computer vision. This study
      presents a comprehensive evaluation of state-of-the-art approaches
      including CSRNet, Mask R-CNN, Faster R-CNN, and YOLOv5, focusing
      on two critical occlusion patterns: full occlusions (umbrellas) and partial
      occlusions (pickets). Our experimental results, evaluated on scenes with
      consistent human presence (GT = 114) but with varying levels of occlu-
      sion, demonstrate distinct performance patterns. Under increasing um-
      brella occlusions (GT: 0,1,10,75) and picket occlusions (GT: 0,1,10,75),
      density-based methods (CSRNet) significantly outperform detection-
      based approaches, maintaining accuracy from 125 to 83 persons. In
      contrast, detection-based methods show substantial degradation: Mask
      R-CNN exhibits a decrease from 74 to 58 persons, Faster R-CNN main-
      tains moderate stability (77 to 69), and YOLOv5 struggles significantly
      (5 to 8). The study reveals that while density estimation approaches
      excel in crowd counting under occlusions, they tend to overestimate in
      clear scenes and underestimate in heavily occluded scenarios. Mean-
      while, instance segmentation methods offer complementary strengths
      in providing detailed instance-level information but suffer from severe
      degradation under heavy occlusions. These findings suggest that future
      research should focus on hybrid approaches that combine density es-
      timation with instance segmentation capabilities for more robust crowd
      counting in complex real-world scenarios.
• S5 - Kathryn Bohanan
  • with Ransom Ward, Prathyu Adari, Riley Bruce, Yugyung Lee
  • "Towards Smarter Cyber Defense: Anomaly Detection with Autoencoders and LLM-Driven Threat Intelligence"
  • Abstract:
    • Network security is increasingly challenged by evolving cyber threats, requiring advanced detection methods to identify and mitigate malicious activities. Our Network Intrusion Detection System monitors traffic in real-time, analyzing data flow before passing it to an autoencoder model, which is trained on the LUFlow dataset. The model classifies network traffic as benign or potentially malicious, helping to detect both known and novel attack patterns. If a threat is detected, the system attempts to shut off the malicious traffic, logs the event, and sends detailed threat information to a Large Language Model (LLM) enhanced with Retrieval-Augmented Generation (RAG). This allows the LLM to provide accurate, context-aware analysis of the flagged data, helping the system administrator or user assess the severity, source, and potential impact of the attack. The LLM analysis will also provide the administrator with mitigation steps, which will help determine further actions such as blocking the source, assessing potential impacts, or verifying false positives. Additionally, users can interact with the LLM in a chatbot-style interface to gain deeper insights into cybersecurity risks, threat trends, and recommended defensive measures. By integrating deep learning, real-time monitoring, and AI-driven intelligence, our NIDS improves threat detection, response time, and decision-making, providing a proactive cybersecurity defense against evolving attacks.
• S6 - Sarah Christian
  • with Hillary McGraw
  • "Mutations in foxg1a Cause Sox10 Olfactory Abnormalities"
  • Abstract:
    • Sensory systems allow animals to interact and adapt to their environment. When errors in development occur, sensory function can be disrupted and can have deleterious impacts on the quality of life for the organism. Mutations in the transcription factor foxg1 result in developmental disorders in humans and in model organisms such as the zebrafish. Research on foxg1 mutations in human patients with Foxg1 Syndrome and in zebrafish shows defects in the olfactory system development. Our work using the foxg1aa266 mutant zebrafish line shows that loss of Foxg1a function results in the specific absence of Sox-positive olfactory neurons. Sox10-positive neurons are one set of cells within the olfactory system that lend to the complexity of the tissue and the ability to respond to a broad range of olfactory stimuli. This defect likely occurs during differentiation of olfactory neurons since neither proliferation nor cell death are affected. The goal of this research is to better understand developmental disorders like Foxg1 Syndrome for their treatment leading to enhanced lives of patients.
• S7 - Jarod Crider
  • with Hee Sup Shin
  • "Electronic Goniometer Designed for Upper Extremities in Dynamic Situations"
  • Abstract:
    • The upper extremities experience constant, complex, and repetitive motions that are the root cause of undue stress and strain within limited joint range. Understanding joint angle movement is a proven method for diagnosing chronic conditions such as arthritis or helping with rehabilitation and injury prevention. Current rehabilitation methods make use of static means of measuring with a trained professional utilizing a goniometer. The system designed takes measurement data from multiple spots with an inertial measurement unit (IMU) and interprets their angle difference using the Madgwick filter. The angle found is to be comparable to static measurements utilizing a goniometer in a short run period. The results show a margin within one degree of expected value, during tests. Ordinarily a magnetometer is used to provide a frame of reference, however due to the nature of the sensor, manual calibration is required but is not possible. To overcome this electromyography (EMG) data is collected and fused with 6-axis IMU data to give dynamic angle reading with long term drift compensated for.
• S8 - Shiva Daneshmehr
  • with Timothy Cox
  • "The ontogeny of craniofacial microsomia in a mouse model"
  • Abstract:
    • Objectives:
      The aim of this study is to characterize the anomalies and postnatal ontogeny of the cranioskeletal features in the small body, small ear (sbse) mouse mutant - one of the few animal models of craniofacial microsomia (CFM).
      Methods:
      Mutant (sbse) and control mice of both sexes were collected at P0, P7, P14, P28, P56, P180, and P356, with up to 12 mice per genotype group. MicroCT imaging of each animal was performed and raw scan data reconstructed and then digitally rendered in 3D. 3D coordinates were then collected for 69 landmarks to facilitate a detailed analysis of cranioskeletal growth differences.
      Results:
      The first localized cranioskeletal size differences in mutants were seen as early as P7, that coincided with premature fusion of the bones of the cranial base in some mutants. By P28, ~95% of mutants exhibited partial or complete fusion of the presphenoid and basisphenoid bones, as well as premature fusion of the zygomatic sutures on one or both sides, which were confirmed histologically. Ontogenic analyses detected midface hypoplasia and cranial doming as early as P21 with a subsequent significant slowdown of skull expansion in mutants after P28. The premaxillary-maxillary suture appeared morphologically abnormal in homozygous sbse mice but did not appear to fuse early, indicating that midfacial hypoplasia (and cranioskeletal asymmetry) in this model may result, at least in part, from constraining forces associated with zygomatic suture fusion and reduced projection secondary to the ablation of the synchondroses.
      Conclusion:
      The sbse mutant mouse recapitulates the major features of CFM. Our ontogenic and histological analysis thus provides clues as to the primary drivers of the dysmorphology in CFM. Ongoing studies of the molecular underpinnings of CFM should aid in the development of more effective surgical management and counseling for patients and families.
• S9 - Rohith Sai Charan Eanugula
  • with Vinay Joshva, Hrishikesh Komaragiri, Sunil Reddy Janapala, Yugyung Lee
  • "AI-Powered Multimodal Marketing Content Generator with Customizable Product Visualization"
  • Abstract:
    • In today’s fast-paced digital marketing landscape, the ability to quickly generate personalized content and visuals is vital for capturing consumer attention and boosting sales. However, creating tailored marketing copy and product imagery typically requires extensive manual effort, technical skill, and time. This project addresses that challenge by developing an AI-Powered Multimodal Marketing Content Generator with Customizable Product Visualization to streamline content creation for e-commerce.
      
      The solution is a full-stack web application built using ReactJS (frontend) and Flask (backend). Users can upload product images to receive AI-generated marketing content via Groq’s LLaMA-3.2 Vision model, engage in interactive chat, and generate customized visuals using Stable Diffusion and ControlNet (Canny Edge Detection). All user sessions, including chat history and image data, are persistently stored and retrieved using MongoDB for continuity and personalization.
      
      Internal evaluations and user testing showed that the system reduced content creation time by over 80%, with users finding the AI-generated descriptions contextually accurate and the customized visuals effective in enhancing product appeal.
      
      This project makes several notable contributions: (1) integrating vision-language AI models for real-time marketing content generation; (2) delivering a seamless, end-to-end interactive content creation tool; (3) providing a scalable system architecture adaptable to various product types and content domains; and (4) demonstrating how multimodal AI can enhance creative workflows, consistency, and efficiency.
      
      In summary, the project presents a user-friendly, intelligent platform that empowers e-commerce professionals and marketers to automate and personalize their digital campaigns showcasing the transformative power of combining multimodal AI with modern full-stack development.
• S10 - Hui Jin
  • with Jayadithya Nalajala, Sai Jahnavi Devabhakthuni, Saniya Pandita, Yugyung Lee
  • "Fall Detection for Elderly People Living Alone Based on Yolov5 and ARKit"
  • Abstract:
    • Falls are a leading cause of injury and hospitalization among the elderly, particularly those
      who live alone without immediate access to assistance. This project addresses this urgent
      concern by developing a real-time fall detection system tailored for elderly individuals
      living independently. The system integrates ARKit, YOLOv5, and Vision Pro. By combining
      ARKit’s spatial perception with YOLOv5’s fast object detection, the system eIectively
      identifies fall events in real time. The algorithm is developed within the iOS environment
      using Swift, Objective-C++, and Python, and utilizes Apple Vision Pro as a high-
      performance input device to capture rich spatial and motion data, enhancing scene
      understanding and detection accuracy. Comparative experiments show that the system
      oIers higher accuracy and faster response than traditional methods. Findings reveal the
      system is especially eIective in detecting slow or unstable movements preceding a fall,
      allowing for earlier intervention. A system architecture diagram is provided to illustrate the
      interaction among components: Vision Pro captures the real-world environment and
      streams data to ARKit for spatial mapping; YOLOv5 processes the visual frames to detect
      people and objects; the fall detection logic integrates all results to determine risk and issue
      alerts.This work not only advances fall detection technology but also oIers a practical,
      socially impactful solution for enhancing the safety and independence of elderly people
      living alone. The algorithm demonstrates strong potential for real-world applications in
      home-based elderly care. Future improvements include optimizing model performance,
      enhancing robustness in complex and dynamic environments, and increasing scalability
      across various living settings.
• S11 - Jason Jordan
  • with Jainil Patel, Krutarth Lad, Harshita Behera, Shu-Ching Chen, Mei-Ling Shyu
  • "LLMout - NFL Multimodal Offensive Strategy Report Generator"
  • Abstract:
    • In the high-stakes world of professional football, accurately classifying offensive plays and predicting success rates provides a critical competitive advantage. This research presents a novel multi-modal deep learning framework that integrates Convolutional Neural Networks (CNNs) for spatial feature extraction and Long Short-Term Memory (LSTM) networks for temporal sequence modeling of NFL player tracking data. Raw positional and player mobility data was collected from NFL Next Gen Stats to predict multiple categories, including offensive formation (OF), receiver alignment (RF), play-action types (PA), run-pass option types (RPO), run concepts (RC), defensive pass coverages (PC), general offensive play types (PT), and whether given play types will allow the offensive teams to successfully move the ball forward a specified distance (OS). For each of these classifiers, the CNN-LSTM model achieved accuracies of 0.56 (OF), 0.41 (RF), 0.80 (PA), 0.89 (RPO), 0.53 (RC), 0.31 (PC), 0.61 (PT), and 0.91 (OS). In addition to the spatio-temporal analysis, the outputs were integrated into a custom Retrieval-Augmented Generation (RAG) vector database, stored in Qdrant, that was augmented with social media and web-based news articles related to NFL team staff and player information. To evaluate the efficacy of the RAG retrieval component, we employed precision@10 as a measure of the fraction of relevant items within the top ten search results, achieving a score of 0.95. This system was paired with a pre-trained large language model (LLM) that teams can utilize through a pre-planned query template to generate useful insights for developing offensive strategies in both pre-game preparations.
• S12 - Abhinav Kochar
  • with Yugyung Lee
  • "PoseCorrect: AI-Driven, Real-Time Posture Correction and Exercise Guidance"
  • Abstract:
    • PoseCorrect is a comprehensive, AI-driven system that provides real-time posture correction and personalized exercise guidance. By combining advanced computer vision, knowledge graph storage, and retrieval-augmented generation (RAG), PoseCorrect addresses the challenges of traditional exercise monitoring and feedback in a holistic manner. At its core, the system leverages ViTPose, a cutting-edge vision transformer-based model, to capture skeletal keypoints from live or recorded video streams. These extracted poses are then refined and structured within a Knowledge Graph, enabling efficient retrieval and comparison across historical user data as well as best-practice reference recordings.
      
      A pivotal component of PoseCorrect is its integration of a Large Language Model (LLM) that taps into the knowledge graph using RAG techniques. By automatically querying relevant past performance records and expert-level exercise examples, the LLM can generate personalized instructions and corrective recommendations in a natural, conversational style. These suggestions are delivered to the user in real time, ensuring that individuals receive immediate, context-aware feedback on their form and movement patterns.
      
      From a user experience perspective, PoseCorrect incorporates a modern, intuitive interface that overlays pose outlines on live video feeds, highlights areas for improvement, and presents step-by-step guidance through a dedicated feedback panel. This dynamic and interactive environment is designed to empower users—whether athletes, patients in rehabilitation, or individuals engaged in general fitness—to correct their form, reduce injury risks, and accelerate progress toward their training goals.
      
      Beyond its direct applications in fitness and physical therapy, PoseCorrect offers broader implications for human-activity recognition and AI-based tutoring systems. By combining explainable AI techniques with structured data storage, the platform paves the way for next-generation so
• S13 - Abhinav Kochar
  • with Karthik Patri, Yugyung Lee
  • "A Quantum-Enhanced Transformer System for Real-Time Intraday Stock Forecasting"
  • Abstract:
    • Intraday trading is characterized by rapid price fluctuations and high volatility, making accurate short-term forecasting crucial for algorithmic traders, market makers, and risk managers. Traditional models like ARIMA, RNNs, and basic LSTMs struggle to capture the non-linear and time-sensitive dynamics inherent in intraday data. To address these limitations, we propose a quantum-enhanced Transformer system designed to generate real-time, hour-ahead stock price predictions. Our framework leverages the Transformer architecture's multi-head attention mechanism for modeling sequential data and introduces day-wise and time-slot-specific segmentation to capture recurring intraday behaviors. Additionally, we integrate an optional quantum neural network extension to explore complex relationships beyond classical models, enhancing the system's ability to handle non-linear patterns in volatile markets.
      
      The data preprocessing pipeline transforms raw OHLC data into feature-specific datasets segmented by time and day, facilitating more granular and context-aware forecasting. By training specialized models for each price feature (Open, High, Low, Close) and time block, the system captures unique intraday patterns while normalizing data with log returns or relative scaling to reduce volatility bias. The modular design supports real-time updates, enabling continuous forecasting with minimal lag.
      
      Through extensive evaluations, our quantum-enhanced models consistently demonstrate lower mean squared errors (MSE) compared to standard Transformers, particularly in predicting volatile price movements like intraday highs. A live trading interface powered by the Yahoo Finance API integrates these forecasts into market operations, displaying both predicted and actual price movements for the Nifty 50 index in real time. This approach bridges the gap in intraday stock forecasting, offering a robust and adaptable solution for real-time financial market analysis.
• S14 - Sai Kamal Makthala
  • with Likhitha Neerati, Sindhu Mukkara, Lalitha Rani Palakaluri
  • "AI Powered Supply Chain Resilience Index(SCRI)"
  • Abstract:
    • Modern supply chains are increasingly vulnerable to disruptions caused by extreme weather, geopolitical tensions, and shifting market dynamics. To address these challenges, we present an AI-Powered Supply Chain Resilience Index (SCRI)—a real-time, intelligent system designed to forecast disruptions and support proactive decision-making in the retail and e-commerce sectors.
      
      Our solution integrates LSTM and Transformer models, enhanced with Low-Rank Adaptation (LoRA), to capture temporal and contextual patterns across customer behavior, logistics, and supplier data. It utilizes real-time weather feeds, economic indicators, and social trend data, which are processed using Retrieval-Augmented Generation (RAG) to generate dynamic risk insights. To ensure transparency and trust, we incorporated explainability frameworks like SHAP and LIME.
      
      The SCRI system features an interactive dashboard built with Streamlit and backend services powered by FastAPI, enabling seamless API integration, user interaction, and live risk updates. The platform supports feedback loops, allowing user input to refine future predictions and improve system adaptability.
      
      Designed for scalability and modular deployment, SCRI empowers retailers, logistics providers, and supply chain managers to anticipate disruptions, evaluate mitigation strategies, and make data-driven decisions. Future enhancements include geospatial visualization, federated learning, and integration of multi-modal AI for broader applicability across complex supply chain networks.
      
      This project demonstrates how AI can transform risk management from reactive to proactive, ensuring continuity, resilience, and customer satisfaction in today’s fast-paced supply chain environments.
• S15 - William McAtee
  • with Keith Buszek
  • "Benzannulated indoles as drug discovery templates. Design and synthesis of libraries inspired by the trikentrin and herbindole natural products and their analogues"
  • Abstract:
    • The trikentrin and herbindole benzannulated natural products and their analogues are attractive templates for diverse small molecule library synthesis and screening initiatives. In this presentation we describe our ongoing efforts to discover and develop novel therapeutics based on these systems, including a next-generation practical and scalable syntheses of these natural products. Although the trikentrins and herbindoles themselves feature only two stereogenic centers and a single functional group (indole NH), the flexibility of the new routes revealed multiple points for diversification and afforded the opportunity to introduce a rich variety of structural elements in each scaffold. This presentation will also describe additional cell-based assays used in the screening studies. Previous work with closely related scaffolds showed that they exhibited unusual levels of activity across multiple therapeutic classes including antimicrobial, anthelmintic, and antiepileptic.
• S16 - Divya Mishra
  • with Apurva Vadlakonda, Vinay Viswanadh Alli, Shivanjali Dyapa, Yugyung Lee
  • "Multimodal Sentiment Intelligence Platform for Dynamic Market Insights"
  • Abstract:
    • Understanding real-time sentiment is increasingly critical in fields like marketing, finance, and customer service. However, traditional sentiment analysis methods—relying solely on text—miss essential emotional cues from facial expressions and visual context. To address these limitations, we present a Multimodal Sentiment Intelligence Platform that integrates both text and visual data to deliver richer, context-aware sentiment analysis.
      The platform uses advanced NLP and Computer Vision models—BERT, GPT-2, YOLOv8, and DeepFace—to extract emotional cues from text and visuals. It is trained on diverse datasets like CMU-MOSEI, IMDb, Amazon Reviews, and FER-2013. We apply feature-level fusion (embedding merge), decision-level fusion (output aggregation), and use RAG with FAISS to incorporate external knowledge for better context.
      Experimental results show the multimodal model outperforms unimodal baselines. The text-only model reached 0.76 accuracy, while our integrated approach achieved 0.80, with gains in F1-score, AUPR, precision, and recall—highlighting the benefit of combining modalities for sentiment analysis.
      The system features a real-time web dashboard (built with Flask) where users input text or media to receive sentiment scores, classifications, and contextual insights. It enables effective customer feedback monitoring, market trend detection, and informed, emotion-aware decisions.
      Challenges remain in sarcasm detection, ambiguous inputs, and demographic bias. Future work will enhance fusion via attention mechanisms, expand datasets, and explore advanced models like CLIP or BLIP-2. Overall, the platform provides a scalable, accurate multimodal sentiment analysis solution with strong real-world utility.
• S17 - An Nguyen
  • with Keith Buszek
  • "Practical and scalable total synthesis of (±)-cis-trikentrin A via 6,7-indole aryne cycloaddition and transition-metal free cross-coupling reactions"
  • Abstract:
    • The trikentrins and herbindoles are the most prominent representatives of an uncommon class of the indole alkaloid natural products in which annulation is present only around the benzene aromatic nucleus. These biologically active and deceptively simple compounds present remarkable synthetic challenges. The trikentrins and the closely related herbindoles have been the subject of numerous synthetic efforts over nearly four decades. The difficulty in their construction is evident by the many creative approaches that have emerged from several laboratories. In this presentation we describe a practical and scalable multi-gram total synthesis of (±)-cis-trikentrin A via the 6,7-indole aryne Diels-Alder cycloaddition and transition-metal free cross-coupling reactions. This much improved third-generation approach from our laboratories takes advantage of our selective C-7 metal-halogen exchange orthogonality in the starting 4,6,7-tribromoindole scaffold, resulting in the formation of the 6,7-indole aryne. The 4,6,7-tribromoindole itself was prepared in just three steps from inexpensive 2-nitroaniline using the Bartoli indole synthesis. The installation of the 4-ethyl substitutent was now accomplished via Wurtz-Fittig coupling instead of the previously employed Pd(0)-catalyzed Negishi cross-coupling reaction. This new approach is also notable for not using any protecting groups for the indole nitrogen throughout the synthesis. Finally, some new biological studies of this natural product will be discussed.
• S18 - Evral Ntsa
  • with John Kevern
  • "Acids effectivenss in hydrodynamic cavitation for methlyl orange degradation in textile wastewater treatment including sludge SEM-EDX characterization"
  • Abstract:
    • This study investigates the effectiveness of three different type of inorganic acids in low pressure orifice based hydrodynamic cavitation (HDC) for the degradation of a model azo dye present in textile wastewater. By creating an acidic environment which is as similar as in textile influent, the system enhances the generation of hydroxyl radicals that are critical for breaking down complex dye molecules. Furthermore, the research compares the effects of this acids on the cavitation process by evaluating the impact degradation on sludges obtained. Analytical techniques, including UV-spectrophotometry and scanning electron microscopic coupled with energy dispersive X-ray spectroscopy were employed to monitor changes in dye absorbance, assess water clarity, and characterize the elemental composition of each acid-based treatment sludge residue post treatment. The findings reveal that the choice of acid markedly impact dye degradation, with one acid demonstrating particularly high removal efficiency and producing water exceptional clarity. However, this enhances performance is accompany by increased corrosion of reactor when containing certain alloys material and which need to be avoided in design as part of the reactor material selection.
• S19 - Amrutha Perumalla
  • with Priyanka Adusumilli, Maheswar Rao Bandi
  • "AI-Powered Job Search Assistant"
  • Abstract:
    • Traditional job matching platforms rely primarily on keyword-based matching, and they tend to yield unsatisfactory outcomes by failing to detect contextual relevance and candidate suitability. To get around these limitations, we suggest an AI-powered job search assistant that enhances job matching, resume parsing, and interview preparation through deep learning, transformer-based NLP models, and FAISS-based vector search. The system utilizes Named Entity Recognition (NER) for resume extraction in structured form, LoRA fine-tuning to produce domain-specific interview questions, and retrieval-augmented generation (RAG) for context-aware job recommendations. FAISS-based Approximate Nearest Neighbor (ANN) search increases the effectiveness of retrieval, BERT embeddings enhance job-role similarity scoring, and GPT-based models streamline resume recommendations. Explainability techniques such as SHAP and LIME facilitate interpretability for AI-driven decisions.
      
      Performance metrics demonstrate significant improvements: job-matching accuracy is 89% (+11% after FAISS optimization), resume parsing F1-score is 85% (+13% after transformer fine-tuning), and interview question BLEU score is 0.82 (+17% after LoRA fine-tuning). Future work involves real-time job retrieval via web scraping, introducing job-matching functionality like skill gap analysis and career path recommendation, and developing a mobile application for convenient cross-platform user experience. This AI-driven platform bridges the gap between job recruiters and candidates, providing individualized and intelligent career guidance.
• S20 - Swagotom Sarkar
  • with Vidit Minda, Megan Hart
  • "Oxidative degradation of Perfluorooctanesulfonic acid (PFOS) in alkaline Ultraviolet/silica granular media: Reaction Mechanism"
  • Abstract:
    • Perfluorooctanesulfonic acid (PFOS) is one form of forever chemicals. Forever chemicals are very persistent in natural environment. Once it was used in different products i.e. firefighting foam, non-stick pans, raincoats, etc. As a result, it is present in water, soil, and air nowadays. Several studies suggested that it is toxic to aquatic and human health. Recently, the US EPA has considered PFOS to be a hazardous substance. Besides, the maximum contaminant level (MCL) for PFOS in drinking water in the USA is 4 ppt. So, it needs to be removed from our environment. According to previous studies, Ultraviolet/Silica granular media (UV/SGM) technology can destroy concentrated PFOS from contaminated water. However, there was no study of how this destruction happens from a mechanistic point of view. This study demonstrated how PFOS is broken down in UV/SGM reactor. To find out the reactive species responsible for PFOS degradation, scavenger experiments were done. This preliminary data suggested that hydroxyl radicals are responsible for PFOS degradation and defluorination. This study will help to make the UV/SGM process faster specifically to break down the sulfate group containing PFAS analytes during field-scale study.
• S21 - Srabonty Soily
  • with Kalyan C. Durbhakula, Deb Chatterjee
  • "Analysis of the Creeping Wave Poles for the Conformal Antennas for (Unmanned Aerial Vehicles ) UAV Applications"
  • Abstract:
    • This study investigates the characteristics of creeping wave poles on dielectric-coated cylindrical structures, essential in ultrawideband conformal antenna design. Creeping waves significantly influence electromagnetic scattering behavior, particularly on curved surfaces. In dielectric-coated cylinders, these waves exhibit distinct propagation behaviors that differ notably from those observed on perfectly conducting cylinders. Creeping wave poles are complex solutions derived from boundary conditions of coated cylindrical geometries. They determine the propagation constant, which governs the phase velocity of creeping waves, and the attenuation constant, which represents the energy loss due to diffraction field leakage, surface impedance, and loss within the dielectric coating material. Unlike planar geometries, dielectric-coated cylindrical structures do not exhibit distinct cutoff frequencies for creeping waves, supporting instead an infinite number of modes. However, the number of low attenuation, Elliott-type, creeping wave modes effectively excited is limited by the coating's dielectric properties and it's thickness. Increasing the dielectric coating thickness profoundly influences the attenuation characteristics, enabling a transition from highly attenuated Watson-type modes to Elliott-type modes with significantly lower attenuation. The comprehensive understanding of creeping wave poles gained through this study will provide the idea of how many higher order creeping wave modes can be considered to design ultrawideband conformal antennas, optimum coating thickness, and optimum radius of the conformal antennas by considering the surface wave losses inside an antenna. This will ultimately help to determine radar cross-section profiles, scattering behavior, and radiation patterns, particularly for conformal antennas mounted on curved aerospace platforms and unmanned aerial vehicles (UAVs).
• S22 - Viswanth Tammana
  • with Dharani Thakkallapally, Yugyung Lee
  • "WanderTales: AI-Powered Travel Stories & Recommendations"
  • Abstract:
    • In an era where travellers seek hyper-personalized and dynamic experiences, traditional itinerary planning tools often fall short struggling to integrate diverse, real-time information and tailor recommendations meaningfully. WanderTales aims to transform the travel planning process by offering an intelligent, immersive, and user-centered solution that adapts to individual preferences while reflecting the cultural and emotional context of the journey. WanderTales leverages cutting-edge AI technologies, including GPT-4, DALL·E, and real-time APIs from Amadeus, OpenWeather, and Google Maps, to create comprehensive, real-time travel itineraries. Built on a Retrieval-Augmented Generation (RAG) framework, the system crafts narrative-driven travel stories that capture the spirit of a destination and the traveler's purpose. In addition to producing tailored plans with recommendations for flights, accommodations, restaurants, and attractions, WanderTales generates AI-powered visuals and videos, offering a cinematic preview of the entire journey. The platform also features an interactive chatbot for conversational trip planning and real-time customization. Preliminary evaluations show that WanderTales significantly enhances user engagement and itinerary relevance by combining real-time contextual data with interactive storytelling. Users experience a deeper connection to their travel plans and find the visual previews particularly valuable in shaping their expectations and decisions. To push personalization even further, future enhancements include AR/VR-based virtual previews, multilingual storytelling, and integrated booking functionalities—positioning WanderTales as a scalable, immersive, and intelligent platform for the next generation of travel experiences.
• S23 - Jennifer Vanderslice
  • with Paul Rulis
  • "Electronic Band Structure Variations in Photothermal Catalytic Materials"
  • Abstract:
    • Novel photothermocatalytic devices based on layered semiconductors have great promise as efficient overall water splitting devices for hydrogen production, which can help make H2 into an economically viable alternative to fossil fuels. The effectiveness of the proposed devices is sensitively influenced by the electronic structure of the different materials and their interfaces. Yet, the electronic band structure properties of many of these materials under thermal expansion is under-explored. In this study we calculate the electronic structures of a select group of sulfide semiconductors and their lattices thermal expansions, focusing on variations in key parameters such as band gap, valence band edges, and conduction band edges.
      
      The chosen materials (Ag2S, CoAsS, FeS2, etc.) were selected for their simple structures, lower electron counts and increasing band gaps compared to each other. By calculating band structures along a universal high-symmetry path that is applicable to all cell symmetries, we determine how temperature-induced asymmetric lattice changes impact the electronic properties of individual materials and the overall thermal stability of the proposed photothermocatalytic device. This approach could provide valuable insights into optimizing materials for efficient photothermocatalytic devices and guide future designs for water splitting applications.
• S24 - Alysse Weigand
  • with Paul Rulis
  • "ab initio Calculation of the Microwave Complex Dielectric Function"
  • Abstract:
    • Density functional theory (DFT) based electronic structure programs (e.g. OLCAO, VASP, CASTEP, etc.) have been used extensively and confidently for the ab initio calculation of the optical properties of materials (e.g., the complex dielectric function (CDF)). While these electronic structure programs can accurately determine the CDF within the visible and ultraviolet range and beyond, they are currently incapable of determining the CDF at lower frequencies. Unlike the DFT calculations of the CDT at optical frequencies, where electrons transition discretely from the valence band to the conduction band in response to a photon by following simple quantum mechanical transition probabilities, at the lower frequencies a time dependent distortion of the electron distribution occurs. Modeling the response of a material to EM radiation below visible frequencies requires the coupling of ionic and electronic degrees of freedom as well as coupling of both to the oscillating EM field. In addition to the electronic, ionic, and dipolar motion, the response of a material to electromagnetic radiation within the microwave frequency range will be time dependent. Combining all aspects presents a complicated theoretical challenge. In fields such as biology, the calculation of a microwave CDF for biomolecules has been determined with the use of linear response theory and molecular dynamics simulations. We present here a methodology for calculating the CDF for bulk materials and nanomaterials in the Microwave frequency range. Our method combines molecular dynamics, DFT, machine learning, and linear response theory to predict the CDF for a variety of materials in the microwave frequency range.
• S25 - Srikanth Yelem
  • with Preetham Goli
  • "Zonal Allocation of Photovoltaic Generators and Battery Storage Units for Resilience Enhancement and Real-Time Volt-Var Control"
  • Abstract:
    • This thesis presents a strategic approach to enhancing the resilience and efficiency of electric distribution networks by integrating fast charging stations, photovoltaic (PV) systems, and battery energy storage (BES). The study focuses on optimizing the placement and capacity of fast charging stations equipped with solar and storage capabilities to support electric vehicle (EV) charging, restore critical loads during outages, and provide voltage stabilization during peak demand. A two-stage methodology is proposed: first, optimal locations for these stations are identified to maximize load restoration and network stability; second, the strategy is validated through real-time co-simulation using OpenDSS, MATLAB, and Typhoon HIL, demonstrating its practical feasibility.
      
      To address the complexity of distribution network resilience, a zonal-based optimization approach is developed. The network is partitioned into distinct zones based on automatic recloser switch locations, allowing strategic allocation of PV distributed generators (PVDGs) and BES units. A Pareto-based multi-objective optimization framework within the fuzzy domain is employed to balance installation costs, power loss reduction, and voltage profile improvements. This ensures an optimal trade-off between PVDG investment and operational efficiency, enhancing overall network performance.
      
      The BES unit sizing is determined using a 24-hour load profile, ensuring comprehensive support across varying demand periods. The proposed optimization methodology is evaluated through simulations on the IEEE 123-bus test system. Results demonstrate that integrating PVDGs and BES units significantly improves network resilience, reliability, and stability, making this strategy a valuable solution for grid operators and planners seeking to enhance grid performance under adverse conditions.
• S26 - Ghadah Almousa
  • with Yugyung Lee
  • "Rumor detection"
  • Abstract:
    • The rapid spread of rumors on social media platforms poses a significant challenge to public trust and the accuracy of information. This paper presents a novel approach to rumor detection through the construction of dynamic graphs, which monitor the evolution of rumors over time and help identify their sources. By analyzing spread patterns as network graphs between users and rumors, we improve the accuracy of rumor classification.
      
      To achieve this, we introduce the Dynamic Graph Attention Partial Differential Equation (DGAPDE) framework for real-time rumor detection in social networks. This framework integrates Graph Convolutional Networks (GCNs), Partial Differential Equations (PDEs), and Attention Mechanisms to model the temporal and spatial dynamics of rumor propagation.
      
      Extensive experiments on the PHEME and Weibo datasets demonstrate that DGAPDE outperforms several state-of-the-art models in rumor detection tasks. Specifically, DGAPDE achieves an average accuracy improvement of 2.2% on PHEME and 1.2% on Weibo, while maintaining high F1 scores across different rumor types. The model effectively handles multi-modal features including text, timestamps, and user interactions, and captures the nonlinear temporal dependencies and node influence patterns critical to rumor diffusion.
      
      The results underscore the importance of modeling the evolving structure of rumor-spreading networks and show that DGAPDE provides a powerful and interpretable framework for understanding and mitigating the impact of misinformation on social media.
• S27 - Saeed Alqarni
  • with Yugyung Lee
  • "SAM-KG: Knowledge Graph-Enhanced Tumor Segmentation with Zero-Shot Foundation Models"
  • Abstract:
    • Accurate tumor segmentation in medical imaging is challenging due to anatomical variability, tumor heterogeneity, and data scarcity. While deep learning models like UNet, Swin-UNETR, and D-LKA Net achieve strong results, they require extensive labeled data and struggle with generalization. Zero-shot models like SAM and MedSAM offer promising alternatives but lack volumetric consistency and structured spatial reasoning. We propose SAM-KG, a knowledge graph-enhanced segmentation framework that integrates multi-modal feature extraction, uncertainty-guided refinement, and anatomical relationship modeling to improve segmentation accuracy and interpretability. Evaluated on the Liver Tumor Segmentation (LiTS) and Medical Segmentation Decathlon (MSD) Pancreas datasets, SAM-KG surpasses state-of-the-art models, achieving superior boundary delineation and volumetric consistency, advancing AI-driven medical image analysis.
• S28 - Udiptaman Das
  • with Yugyung Lee
  • "KARMA: A Knowledge Graph Framework for Explainable and Personalized AI in Healthcare"
  • Abstract:
    • The growing complexity and scale of healthcare data—from structured clinical records to unstructured notes and genomics—pose significant challenges for effective analysis, integration, and clinical decision-making. While AI models such as transformers and biomedical language models (e.g., BioBERT, SciBERT) have enhanced diagnostics and predictive capabilities, they often lack explainability, semantic structure, and interoperability with clinical standards.
      
      We present KARMA (Knowledge Application for Reports in Medical Analysis), a novel framework that leverages large language models (LLMs) and medical ontologies to construct patient-centric, ontology-enriched knowledge graphs. Using Gemini Flash 2.0, each step of the KG pipeline—from entity and attribute extraction to schema generation and RDF structuring—is LLM-guided. Standard ontologies like SNOMED CT and ICD-11 are integrated for semantic alignment, enabling scalable and interoperable graph construction.
      
      KARMA further incorporates AI-driven community detection (Louvain modularity) to identify clinically meaningful patient subgroups and applies semantic reasoning (SPARQL, SWRL) to support personalized recommendations. A dynamic query system enables clinicians to explore graphs interactively using natural language inputs.
      
      To enhance graph quality, a second-pass refinement using GPT-o1 and Grok restructures the entire KG by normalizing entity-attribute-sub attribute hierarchies and refining relationships and schema. This step improves semantic precision, reduces hallucinations, and increases trust in LLM-driven clinical reasoning. Evaluation will include metrics such as ontology coverage, semantic consistency, hallucination rate, and information gain.
• S29 - Jenna DeCata
  • with Maria Spletter
  • "Multiple isoforms of RNA-binding protein Bruno1 are required during indirect flight muscle development in Drosophila"
  • Abstract:
    • Indirect flight muscles (IFMs) of Drosophila undergo a switch in fiber-type specific alternative splicing that is essential for flight behavior. This switch is regulated by RNA-binding proteins (RBPs), especially Bruno1 (Bru1). Bru1 is a conserved member of the CELF family, which are important regulators of developmental alternative splicing in vertebrate striated muscle. CELF activity is misregulated in patients with Myotonic Dystrophy Type I, notably reverting to embryonic splicing patterns. In flies, Bru1 has a role in translation repression during embryo formation. Recently, Bru1was identified as a muscle fiber-type specific splicing factor necessary for a developmental transition to IFM-specific splice isoforms. Bru1 promotes early cytoskeletal rearrangements enabling myofibrillogenesis and maturation of the sarcomeres and contractility. Bru1 is alternatively spliced, but isoform-specific CELF functions are not reported in flies or vertebrates. Here we test if Bru1 isoforms play distinct roles in myogenesis. We show that an isoform-specific CRISPR mutant affecting long Bru1 isoforms develops normally but has a hypercontraction phenotype. Using Gal4-UAS to control when and where specific Bru1 isoforms are expressed, we test the function of long (isoform B), middle (isoform A), and short (isoform D) isoforms. Our results show distinct localization patterns for Bru1 isoforms in myogenesis. We show that overexpression of Bru1-isoB results in a strong muscle detachment phenotype, while Bru1-isoA overexpression disrupted sarcomere structure. Bru1-isoD overexpression resulted in only minor muscle defects. We further tested isoform-specific rescue ability and find that a combination of Bru1-isoA and Bru1-isoD are required to rescue IFM development. These differences suggest a requirement for distinct Bru1 isoforms during muscle development and are the first demonstration of CELF-family isoform-specific function in myogenesis.
• S30 - Chuheng Xiao
  • with Yugyung Lee
  • "Grocery Tracker & Recommendation Generator AI for Smarter Grocery Choices and Meal Planning"
  • Abstract:
    • Grocery Tracker & Recommendation Generator is driven by a clear mission: to reduce food waste, improve meal planning, and empower users to make healthier, more efficient grocery decisions with minimal effort. In a world where busy lifestyles often lead to poor food organization and uninformed choices, our goal is to create an intelligent, user-friendly system that bridges the gap between what people have and what they could do with it.
      At the heart of the app is a personalized experience. Users can track their groceries effortlessly by simply uploading photos, receive AI-powered meal and grocery suggestions tailored to their health goals and dietary needs, and generate recipes based solely on what’s in their pantry. Whether the user is focused on saving time, reducing waste, or eating better, our app adapts to their needs.
      Unlike existing tools that treat inventory, planning, and cooking as separate workflows, we unify them. The system leverages computer vision and natural language processing in a single, seamless Streamlit interface. Google’s Gemini 1.5 model powers image recognition, while a recommendation engine curates smart suggestions. The platform is designed with future enhancements in mind, such as visual recipe instructions, nutritional insights, mobile optimization, and scalable backend support via MongoDB.
      By blending personalization, simplicity, and smart automation, Grocery Tracker & Recommendation Generator redefines how people engage with their food—from pantry to plate.
• S31 - M. Atikur Rahman
  • with Zhonghua Peng, Masud Chowdhury
  • "Controlling Growth of 2D Perovskite Through Novel Spacer Cation Engineering for Optimized Sensing"
  • No abstract provided